Summary
Two tools, and the shape of the loop is stated rather than left to whoever wires it up.
recall belongs at the start of a turn and advanced_augmentation after the response has been drafted, which puts the read before the reasoning and the write after it. That ordering is the whole integration: an agent calling neither has no memory, and one calling them in the wrong order stores context it has not used yet. What gets kept is described as durable facts and preferences rather than a transcript, so the store is meant to hold conclusions rather than history.
What it is
A memory server with two tools — retrieve the memories relevant to a query, and store durable facts and preferences from an exchange.
What you get
- Recall retrieves the memories relevant to a query, meant to run at the start of a user turn to bring back prior context, preferences and facts.
- advanced_augmentation stores durable facts and preferences from the user's message and the assistant's response, meant to run after replying.
Requirements
Nothing — no account, no key.
Setup effort
One command plus a key — npx -y @memorilabs/memori-mcp, then supply credentials